Fluid Antenna Arrays Overview

• Fluid antenna arrays are reconfigurable-aperture architectures that adjust physical or electromagnetic configurations to optimize wireless performance.
• They exploit additional spatial degrees of freedom, using movable elements or software-controlled radiation states for applications such as DOA estimation, AirComp, and ISAC.
• Advanced optimization methods and practical hardware implementations like ER-FAS and meta-fluid antennas deliver improved efficiency and robustness compared to fixed arrays.

Fluid antenna (FA) arrays are reconfigurable-aperture antenna architectures in which the effective radiative configuration is not fixed in space or electromagnetic state. In the literature, this reconfigurability appears in several forms: an antenna can switch among preset locations inside a predefined region, antenna elements can be repositioned along a constrained support and represented by an antenna position vector (APV), or each element’s electromagnetic radiation state can be software-controlled without physical displacement. The common feature is the introduction of additional spatial or electromagnetic degrees of freedom relative to fixed-position antenna (FPA) arrays, with reported applications spanning multiple access, over-the-air computation (AirComp), direction-of-arrival (DOA) estimation, integrated sensing and communication (ISAC), multiuser MIMO, cell-free massive MIMO, near-field physical-layer security, and wideband 5G NR OFDM (Wong et al., 2020, Xu et al., 27 Nov 2025, Wang et al., 27 Feb 2025, Liu et al., 15 Sep 2025).

1. Conceptual scope and architectural variants

The early FA formulation treated a fluid antenna as a device that can “appear instantly at one of $N$ preset locations in a predefined space,” primarily to exploit small-scale fading diversity in a compact device with a single RF chain. In that formulation, Fluid Antenna Multiple Access (FAMA) selects the port with the highest signal-to-interference ratio (SIR), using interference deep fades rather than sophisticated DSP to create favorable reception conditions (Wong et al., 2020).

Subsequent work broadened the notion of FA arrays from single-port selection to multi-element, position-reconfigurable arrays. In AirComp and related uplink settings, the AP is equipped with $N$ FAs positioned along a line segment of length $L$, and the APV $\boldsymbol{x}=[x_1,\ldots,x_N]^T$ becomes an explicit optimization variable. In these models, the salient contrast with FPA arrays is that the array geometry itself participates in the transceiver design (Zhang et al., 2023, Pakravan et al., 22 Apr 2025).

The architectural scope expanded further in two directions. First, electromagnetically reconfigurable fluid antenna systems (ER-FAS) were proposed to control each element’s radiation pattern directly rather than only its position. ER-FAS thus moves beyond spatial reconfigurability and introduces element-level electromagnetic reconfigurability, with software-controlled fluidics altering director and reflector states of each element (Wang et al., 27 Feb 2025). Second, meta-fluid antennas realized “true electromagnetic fluidity” through electronically reconfigurable meta-atom slot elements, energized by a single RF chain through a substrate-integrated waveguide, again without mechanical motion (Liu et al., 15 Sep 2025). This makes “fluid antenna array” an umbrella term for a broader reconfigurable-aperture class rather than a single mechanical implementation (Xu et al., 27 Nov 2025).

2. Geometry, aperture constraints, and analytical limits

A recurring FA array model uses one-dimensional continuous placement under hard physical constraints: $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$ where $L_0$ enforces minimum spacing to avoid mutual coupling. In several communication designs, these geometric constraints are coupled to movement budgets or total energy budgets, so the feasible set is not merely kinematic but resource-constrained (Zhang et al., 2023, Pakravan et al., 22 Apr 2025, Li et al., 5 Dec 2025).

Finite-aperture analysis has made the geometry-performance relation explicit. For a linear array with normalized aperture, one result gives

$$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$

so the Fisher information is governed by the geometric variance of the port locations. This unifies conventional and reconfigurable arrays within the same CRB expression and yields the explicit design rule that maximizing geometric variance minimizes CRB, while simultaneously exposing a precision–ambiguity trade-off because widely spread ports can increase sidelobe- or grating-lobe-induced ambiguity (Zhang et al., 26 Jan 2026).

For planar finite-aperture FAAs, the corresponding estimation structure is governed by a $2\times 2$ geometric inertia matrix. Its determinant and eigenstructure fully capture the role of port placement in joint elevation-azimuth estimation precision, and both trace and determinant are invariant to the azimuth look direction. The same planar analysis shows that maximizing the geometric determinant pushes ports toward the aperture boundary, again revealing an intrinsic precision–ambiguity trade-off (Zhang et al., 21 May 2026).

Minimum-spacing analysis also differs sharply between linear and planar placements. For random linear FAA placement inside a fixed aperture $W_{\max}$, the expected minimum spacing is

$$\mathbb{E}[\Delta_{\min}] = \frac{W_{\max}}{M^2-1},$$

which decays quadratically with $N$0. By contrast, for uniform random placement over a rectangular planar aperture, the minimum inter-port distance follows an approximate Rayleigh law with mean proportional to $N$1, a substantially more favorable packing behavior than the linear case (Zhang et al., 26 Jan 2026, Zhang et al., 21 May 2026).

Continuous position design has also been shown to alter classical sparse-array DOF limits. For sparse FAS design over a region $N$2, one universal dual DOF bound is

$N$3

The same analysis states that FAS-optimized positions can approach the geometric bound and that, for $N$4 sources, the CRB scales as $N$5, rather than saturating under grid-constrained designs (Wu et al., 19 May 2026).

3. Optimization formulations and algorithmic patterns

Most FA-array design problems are formulated as non-convex joint optimizations in which beamforming, power or port activation, and geometry are tightly coupled. In AirComp, the canonical objective is mean-squared error (MSE) minimization over user transmit coefficients, AP receive beamforming, and the APV. One representative formulation minimizes $N$6 subject to per-user power, spatial-boundary, minimum-spacing, and total-energy constraints that may include FA movement energy (Zhang et al., 2023, Li et al., 5 Dec 2025).

The dominant solution pattern is alternating optimization or block coordinate descent. In the hardware-impaired AirComp setting, the $N$7-update admits the closed form

$N$8

with $N$9 obtained by bisection, the receive beamformer update is the least-squares/MMSE-type solution

$L$0

and the APV update is handled by projected gradient descent with finite-difference gradients, Armijo backtracking, and projection onto the movement/spacing-feasible region. The resulting BCD method monotonically decreases the MSE and converges to a stationary point (Li et al., 5 Dec 2025).

Robust AirComp under angle-of-arrival uncertainty adopts a similar three-block structure, but the APV update is performed with a BFGS quasi-Newton method plus a logarithmic barrier function

$L$1

making the sensitivity of the channel error term to $L$2 explicit (Pakravan et al., 22 Apr 2025). Earlier FA-AirComp work used alternating optimization with QCQP transmit updates, a convex least-squares decoding-vector update, and a primal-dual interior-point method for the APV (Zhang et al., 2023).

Other FA applications instantiate the same general pattern with different block structures. UAV-enabled ISAC introduces a three-timescale design in which UAV trajectory is optimized on a long timescale, FA positions on a medium timescale with update period $L$3, and beamforming every slot; its alternating-optimization solver combines SDR, fractional programming, second-order Taylor-based convexification, and SCA (Liu et al., 25 Sep 2025). FA-assisted MU-MIMO with decentralized baseband processing partitions the FA array into clusters and uses a decentralized BCA framework based on matrix fractional programming and majorization-minimization, reducing computational time by over $L$4 relative to centralized architectures with negligible weighted-sum-rate loss (Liao et al., 8 May 2025).

When the problem is sparse port selection rather than continuous placement, the optimization viewpoint changes. Flexible beam synthesis with a two-dimensional planar FAA is formulated as

$L$5

and is solved by a tailored OMP-like compressive-sensing procedure accelerated by FFT, together with an iterative FFT-based phase-retrieval method requiring only one FFT and one inverse FFT per iteration (Xu et al., 27 Nov 2025). In user-centric cell-free massive MIMO, FA reconfigurability during pilot transmission is embedded in a generalized LMMSE estimator and coupled to a local NMSE-descent port-selection strategy initialized by a log-det correlation criterion (Olyaee et al., 28 Dec 2025).

4. Estimation, sensing, and beam synthesis

FA arrays have been particularly prominent in DOA estimation because mobility can synthesize virtual apertures or difference arrays far larger than the number of physical elements. Under time-constrained mobility, aligned received signals (ARS) use a fully movable structure in which $L$6 FAs and $L$7 movements generate a virtual ULA with $L$8 positions, whereas non-aligned received signals (NARS) use one fixed reference antenna plus $L$9 movable FAs to generate a virtual difference array with $\boldsymbol{x}=[x_1,\ldots,x_N]^T$0 unique lags, $\boldsymbol{x}=[x_1,\ldots,x_N]^T$1 (Xu et al., 14 Aug 2025).

The corresponding estimation algorithms are TMRLS-MUSIC for ARS and TMR-MUSIC for NARS. Both methods address the large-virtual-array, low-snapshot regime by enforcing Toeplitz structure on the covariance, and both use Nyström approximation to reduce subspace-extraction complexity. Reported results state that the ARS and NARS FA schemes can resolve up to $\boldsymbol{x}=[x_1,\ldots,x_N]^T$2 and $\boldsymbol{x}=[x_1,\ldots,x_N]^T$3 sources, respectively, and that adding a single FA movement can be more beneficial than increasing the number of collected snapshots by an order of magnitude (Xu et al., 14 Aug 2025).

A separate sparse-FAS line develops bespoke aligned and misaligned array structures with closed-form, LoS-centric DOA estimation. There, path-number detection is based on an eigenvalue-ratio test, followed by polynomial root-finding rather than a spectral grid search. Reported simulations show estimation of up to $\boldsymbol{x}=[x_1,\ldots,x_N]^T$4 LoS sources with only $\boldsymbol{x}=[x_1,\ldots,x_N]^T$5 FAs in the aligned case and up to $\boldsymbol{x}=[x_1,\ldots,x_N]^T$6 LoS sources with $\boldsymbol{x}=[x_1,\ldots,x_N]^T$7 antennas in the misaligned case (Xu et al., 14 Aug 2025).

Continuous sparse position design pushes this further. The two-stage FAS-MUSIC procedure first applies coarray MUSIC for disambiguation and then performs a local ML refinement over the full physical FAA aperture. Extensive simulations show that FAS-MUSIC achieves $\boldsymbol{x}=[x_1,\ldots,x_N]^T$8 lower RMSE than ULA-MUSIC and that FAS with $\boldsymbol{x}=[x_1,\ldots,x_N]^T$9 antennas outperforms MRA with $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$0 antennas, results tied directly to the large-aperture continuous design rather than grid-constrained sparse geometry (Wu et al., 19 May 2026).

FA arrays have also been integrated into ISAC sensing objectives. In UAV-enabled ISAC, transmit and receive FA positions are optimized jointly with beamforming and trajectory to maximize downlink sum rate while minimizing the Cramér–Rao bound for target elevation-angle estimation. For the receive FA array, the analysis shows that maximizing total sum of squares of the receive coordinates is achieved by spreading elements at the array boundaries, thereby maximizing aperture (Liu et al., 25 Sep 2025). In flexible beamforming with two-dimensional planar FAAs, the reported effect of port selection is improved beam-pattern reconstruction accuracy relative to conventional fixed-array architectures, including more accurate synthesis of both narrow and broad beams (Xu et al., 27 Nov 2025).

5. Communication, AirComp, and networked FA systems

AirComp is one of the most developed FA-array communication applications. In the original FA-enhanced AirComp model, joint optimization of user transmit equalization, AP decoding, and APV significantly outperformed FPA arrays in MSE, with the performance gap increasing as the number of antennas $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$1 and the number of users $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$2 increased (Zhang et al., 2023). Robust AirComp under AoA uncertainty added a channel-error term depending on $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$3, showing that FA positioning directly affects robustness to CSI error, and numerical results established the necessity of uncertainty-aware resource allocation (Pakravan et al., 22 Apr 2025). Hardware-impaired AirComp then incorporated transceiver distortion noise with level parameter $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$4 and FA movement energy consumption into the same optimization, reporting reliable convergence within a few dozen iterations even with $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$5 and $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$6, lower MSE than FPA arrays for all tested $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$7, and more graceful degradation as $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$8 increased (Li et al., 5 Dec 2025). A complementary copula-based analysis derived a closed-form CDF for AirComp MSE under spatially correlated FA ports via the Gumbel copula and showed that FA deployment substantially reduces the MSE relative to fixed antennas, although the gain decreases as spatial correlation becomes stronger (Pakravan et al., 5 May 2026).

In wideband communications, FA systems were integrated into 5G NR OFDM through a port-selection matrix and an adaptive modulation and coding scheme driven by achievable-rate and BICM-capacity calculations. Link-level simulations under 3GPP TDL channels reported up to $$0 \leq x_1 < \cdots < x_N \leq L, \qquad x_n-x_{n-1}\geq L_0,$$9 dB gain in SNR at BLER $L_0$0, and for $L_0$1 MHz bandwidth with $L_0$2 and $L_0$3 ports, throughput reached $L_0$4 Mbit/s versus $L_0$5 Mbit/s for FPA (Hong et al., 7 Mar 2025).

FA arrays have also been inserted into larger network architectures. In large-scale cellular networks, a circular multi-FA array at the UE is paired with skip-enabled sequential LMMSE channel estimation and a two-stage SINR-based port-selection rule; the analysis highlights the trade-off between diversity gain and channel-estimation quality under limited coherence time (Skouroumounis et al., 2022). In user-centric cell-free massive MIMO, FA-equipped APs use generalized LMMSE channel estimation with dynamic pilot-phase port activation, distributed port selection to reduce estimation NMSE, and alternating optimization of data-phase port configurations to maximize uplink sum spectral efficiency (Olyaee et al., 28 Dec 2025). In FA-assisted MU-MIMO, decentralized baseband processing partitions large FA arrays into clusters for parallel optimization and yields substantial runtime reduction with negligible weighted-sum-rate loss (Liao et al., 8 May 2025).

Multiple-access formulations retain the original FA emphasis on single-RF-chain spatial selection. FAMA derived outage-probability bounds and an outage-capacity lower bound showing that hundreds of users can, in principle, be supported in a few wavelengths of space with only one fluid antenna at each user (Wong et al., 2020). Meta-fluid antenna multiple access extends this to multi-activation with a single RF chain, no CSI, and pattern selection based only on SIR measurement; the reported prototype optimizes SIR within a $L_0$6 timeframe and achieves SINR $L_0$7 dB for all users within $L_0$8 pattern switches (Liu et al., 15 Sep 2025).

6. Hardware realizations, near-field operation, and design tensions

FA-array research now includes fabricated hardware rather than only abstract array models. ER-FAS prototypes use Galinstan liquid metal alloy for single elements, fluid silver paste for larger arrays, software-controlled syringe pumps, and a 12-element 1D array with jointly optimized phase shifts and radiation states. Experimental reports include impedance bandwidth below $L_0$9 dB from $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$0 to $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$1 GHz, simulated gain improvements up to $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$2 dB at $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$3, measured gain enhancements of $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$4 dB at $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$5 and $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$6 dB at $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$7, up to $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$8 bps/Hz spectral-efficiency improvement over conventional arrays, and indoor SDR trials showing $$\mathrm{CRB}(\theta)=\frac{1}{8\pi^2 T\,\mathrm{SNR}\,\sin^2\theta\,\mathcal{L}_{\text{geo}}(\boldsymbol{p})}, \qquad \mathcal{L}_{\text{geo}}(\boldsymbol{p})=\sum_{m=1}^{M}(p_m-\bar p)^2,$$9–$2\times 2$0 dB received-power gains together with lower BER for 4-QAM (Wang et al., 27 Feb 2025).

Meta-fluid antenna hardware demonstrates a distinct implementation path. The prototype uses $2\times 2$1 meta-atoms arranged in $2\times 2$2, each controlled by four PIN diodes and fed through an SIW from a single RF chain. The reported switching time is below $2\times 2$3, operation is at $2\times 2$4 GHz, and theoretical, full-wave EM, and measured results are described as closely aligned (Liu et al., 15 Sep 2025).

Near-field FA-MIMO security work adds another practical distinction: FA benefits are not uniform across aperture regimes. In compact arrays, beamforming alone can provide almost zero secrecy rate when Eve is closer to Alice than Bob, whereas joint beamforming and artificial noise with FA port selection significantly improves secrecy. In large arrays, by contrast, the need for artificial noise diminishes because near-field focusing itself can nearly nullify Eve’s channel; FA port selection nevertheless continues to provide measurable secrecy-rate gain over static FPA arrays (Zhang et al., 2 Dec 2025).

Several recurring misconceptions are therefore contradicted by the technical record. FA gains are not solely a consequence of “more ports”: strong spatial correlation can erode selection diversity, training overhead can offset diversity gains under limited coherence, and finite-aperture designs face explicit precision–ambiguity trade-offs (Pakravan et al., 5 May 2026, Skouroumounis et al., 2022, Zhang et al., 26 Jan 2026). Nor are FA arrays synonymous with mechanical motion: ER-FAS and meta-fluid antennas obtain reconfigurability through electromagnetic state control rather than bulk displacement (Wang et al., 27 Feb 2025, Liu et al., 15 Sep 2025). A plausible implication is that FA-array research is now less a single array model than a family of reconfigurable-aperture design paradigms whose performance hinges on how geometry diversity, fading diversity, training resources, and hardware constraints are jointly managed.